Modelling acute HIV infection using longitudinally measured biomarker data including informative drop-out.

Abstract

Background. Numerous methods have been developed to model longitudinal data. In HIV/AIDS studies, HIV markers, CD4+ count and viral load are measured over time. Informative drop-out and the lower detection limit of viral load assays can bias the results and influence assumptions of the models. Objective The objective of this thesis is to describe the evolution of HIV markers in an HIV-1 subtype C acutely infected cohort of women from the CAPRISA 002: Acute Infection Study in Durban, South Africa. They were HIV treatment naive. Methods. Various linear mixed models were fitted to both CD4+ count and viral load, adjusting for repeated measurements, as well as including intercept and slope as random effects. The rate of change in each of the HIV markers was assessed using weeks post infection as both a linear effect and piecewise linear effects. Left-censoring of viral load was explored to account for missing data resulting from undetectable measurements falling below the lower detection limit of the assay. Informative drop- out was addressed by using a method of joint modelling in which a longitudinal and survival model were jointly linked using a latent Gaussian process. The progression of HIV markers were described and the effectiveness and usefulness of each modelling procedure was evaluated. Results. 62 women were followed for a median of 29 months post infection (IQR 20-39). Viral load increased sharply by 2.6 log copies/ml per week in the first 2 weeks of infection and decreased by 0.4 log copies/ml per week the next fortnight. It decreased at a slower rate thereafter. Similarly CD4+ count fell in the first 2 weeks by 4.4 square root cells/ul per week then recovered slightly only to decrease again. Left-censoring was unnecessary in this acute infection cohort as few viral load measures were below the detection limit and provided no improvement on model fit. Conclusion. Piecewise linear effects proved to be useful in quantifying the degree at which the HIV markers progress during the first few weeks of HIV infection, whereas modelling time as a linear effect was not very meaningful. Modelling HIV markers jointly with informative drop-out is shown to be necessary to account for the missing data incurred from participants leaving the study to initiate ARV treatment. In ignoring this drop-out, CD4+ count is estimated to be higher than what it actually is.